Using a hierarchical properties ranking with AHP for the ranking of electoral systems

Transcription

1 Università di Pisa Dipartimento di Informatica Technical Report: TR Using a hierarchical properties ranking with AHP for the ranking of electoral systems Lorenzo Cioni lcioni@di.unipi.it September 29, 2008 ADDRESS: Largo B. Pontecorvo 3, Pisa, Italy. TEL: FAX:

2

3 Using a hierarchical properties ranking with AHP for the ranking of electoral systems Lorenzo Cioni lcioni@di.unipi.it September 29, 2008 Department of Computer Science, University of Pisa lcioni@di.unipi.it Abstract Electoral systems are complex entities composed of a set of phases that form a process to which performance parameters can be associated. One of the key points of every electoral system is represented by the electoral formula that can be characterized by a wide spectrum of properties that, according to Arrow s Impossibility Theorem and other theoretical results, cannot be all satisfied at the same time. Starting from these basic results the aim of this paper is to examine such properties within a hierarchical framework, based on Analytic Hierarchy Process proposed by T. L. Saaty (Saaty (1980)), performing pairwise comparisons at various levels of a hierarchy so to get a global ranking of such properties. Since any real electoral system is known to satisfy some of such properties but not others it should be possible, in this way, to get a ranking of the electoral systems according also to the political goals both of the voters and the candidates. In this way it should be possible to estimate the relative importance of each property with respect to the final ranking of every electoral formula. 1 Introduction The present paper contains both the description of a ranking method and some applications of that ranking method to the properties we wish our voting systems to satisfy. Our aim is to investigate whether through a hierarchic ranking of properties we can devise a ranking of electoral methods or even an electoral method. Extended version of the paper Ranking electoral systems through hierarchical properties ranking to appear on a special issue of Homo Oeconomicus devoted to the quantitative assessment of electoral systems. 1

4 The paper is structured as follows. After a brief discussion of the basic motivations of the paper and a very short description of the mathematics of the ranking method we propose some notes on electoral systems and list, with some comments, the properties we wish the various electoral systems to satisfy. Afterwards we present a simple ranking example and propose it as a voting method. The next step is the application of the ranking method to more complex cases so to get a certain number of ranked orderings. The last step is the association between some orderings and some more or less classical voting methods. The paper closes with some remarks and plans for future works. 2 The basic motivations The basic motivations of this paper are the definition of a ranking method and tool to perform global or social rankings and its application at two paradigmatic cases: as a voting method (see section 6) and as a ranking tool of the properties of a set of voting methods (see sections 7 and 8) for the selection of a perfect voting system. In our context we say that a voting system is perfect if it fits at the best a set of properties that those who must use it (or a pool of appointed experts) think are relevant and must be satisfied. The ranking method and tool we chose is the Analytic Hierarchy Process (that we briefly introduce in section 3) since we think it is the proper tool for such a task and that it can potentially avoid the theoretical hindrances we list in section 4 and, at he same time, allow the fulfillment of the proper sets of properties we examine in section 5. By its very nature, the Analytic Hierarchy Process requires that each involved actor, either in isolation or in co-operation with the others 1, performs the proper rankings in real cases. Since, however, our primary goal was its presentation and the demonstration of its use and usefulness in performing such rankings (see Saaty (1980) and di Cortona et al. (1999)) we did not use a sample of real actors that ranked real cases. On the contrary the rankings have been executed by a single person (or two at the most) and many of them have been performed having in mind more the formal aspects of the Analytic Hierarchy Process or the need to get consistent matrices than any deep comparison among the involved properties. In this way our aim was to show the formal aspects of the tool and, at the same time, to test it and show its usefulness. This paper therefore presents the Analytic Hierarchy Process and tests it on toy examples to show its potentialities but, obviously, it must be thoroughly tested in real cases with real actors that must perform real choices so to be fully validated (see section 10). 1 This depends, as we show in section 3, on the structure of the hierarchy that is associated to the ranking problem. 2

5 3 The mathematical tool The Analytic Hierarchy Process is both a method and a tool developed by T. L. Saaty (Saaty (1980)) and used by himself and others in many fields (see for instance Saaty and Kearns (1985) and Bhushan and Rai (2004).) 2. It represents a useful investigation tool in all cases we have to rank n alternatives depending on their order of importance or preference (with respect to some actors 3 and to a general or main goal) on the basis of qualitative valuations expressed using numerical values on a ratio scale. The method starts with an analysis phase that allows the identification of a set of elements and the definition of a hierarchy among these elements under the form of a rooted hierarchy 4. At the root (level l = 0) we have the main goal, in many cases of political nature, at level 1 we may have the actors, at level 2 the policies, at level 3 the criteria and, last but not least, at the level of the leaves we have the alternatives 5. In Figure 6 1 we show a somewhat simplified complete hierarchy with a main goal (MG), three actors (ac1, ac2 and ac3), four criteria (the cri) and three alternatives (A, B and C). Our example rooted hierarchy has therefore a maximum depth 7 d = 3. Given any level i [0, d 1], if we want to evaluate the importance of the elements at level i+1 with respect to those at level i we can build m matrices of size n n where n is the number of elements at level i +1 and m is the number of elements at level i. In case of Figure 1 we have one 3 3 matrix to weight the importance of the actors with respect to the main goal, three 4 4 matrices to weight the importance of the criteria with respect to each of the actors and four 2 The topic is really complex and wide. It is obvious that in this section we cannot scratch but the very surface. Anyway Saaty (1980), though hard to find, is a good starting point. We note that the method has also been criticized in many papers, among which e Costa and Vansnick (16 June 2008), where the authors stress how the method does not satisfy a fundamental measurement condition and how this makes its use as a decision support tool really problematic, e Costa and Vansnick (16 June 2008), page 14. In our opinion this serious criticism affects only marginally our work where we are mainly interested in getting global orderings (with possible ties) of a set of alternatives. The criticism is however really serious and we have inserted it within the open theoretical problems (see section 10). 3 In this paper we use the terms actor and voter as synonyms though with the former we stress a role within Analytic Hierarchy Process whereas with the latter we stress a role within a voting system. 4 We have a hierarchy with a root at level l = 0 and a set of elements at the deepest level from the root that we call leaves and that contains the alternatives we want to rank with respect to the root. The hierarchy does not contain any cycle because the arcs are covered either from the root to the leaves (analysis phase) or from the leaves to the root (synthesis phase). 5 Any of these levels may be missing but, on the other hand, more levels can be added if this is required by the problem at hand. The minimal number of levels is three: the main goal at level l = 0 and two more levels: actors and alternatives. It should be obvious that a minimum number of two actors is needed in order the overall process to have any meaning. 6 A hierarchy is defined complete if all the elements at two contiguous levels are connected by exactly one arc (so that if they have respectively n 1 and n 2 elements between the two levels we have n 1 n 2 arcs) otherwise it is termed as incomplete. In this paper we are going to consider only hierarchies of the first type. 7 With the term depth we define the number of arcs from an element of the hierarchy to the root along the shortest path. 3

6 Figure 1: Example of a complete hierarchy 3 3 matrices to weight the alternatives with respect to each of the criteria. All this represents what Saaty calls the analysis phase. Such phase is carried out by the actors that have a common goal and that, either individually or in co-operation, evaluate the matrices of the pairwise comparisons, matrices that must be properly merged (see Saaty (1980) for further details). Between each pair of consecutive levels i and i + 1 each matrix is evaluated performing pairwise comparisons between the elements of level i + 1 with regard to those at level i. If we call A one of those matrices we have that its elements 8 a ij (with i, j = 1,...,n) assume positive values from an a priori defined scale and satisfy the following conditions: 1. a ii = 1 2. a ji = 1 a ij If matrix A satisfies such properties it is called positive reciprocal. We note that a ij can assume a value from the following scale of values (Saaty (1980)): 1 to denote equal importance, 3 to denote a weak importance of one over the other, 5 to denote essential or strong importance of one over the other, 7 to denote very strong or demonstrated importance of one over the other, 9 to denote absolute importance of one over the other, 2, 4, 6 and 8 to denote intermediate values whereas a ji assumes the reciprocal value (or vice versa). At this point (see Figure 1) we have to switch to the synthesis phase 9 whose aim is the definition of a normalized vector of priorities of the three alternatives with respect to the main goal. The calculation of such vector turns into a series of eigenvalue/eigenvector problems. To see how this can hold we need some preliminary steps. 8 Element a ij represents the relative importance of element i with respect to element j so that a ii represents the relative importance of an element with respect to itself and it is therefore equal to 1. 9 The synthesis phase is a purely computational phase whose aim is the evaluation of a vector of priorities with the highest accuracy. 4

7 Once the matrices have been defined we have to define for each of them a normalized vector 10 w of weights w i [0, 1]. Such weights are obviously not known in advance otherwise we could write: a ij = w i w j (1) with i, j = 1,...,n. For the moment let us suppose we live in such an ideal world so that those weights are known. From (1) we can get: or (through simple algebra): a ij w j w i = 1 (2) n a ij w j = nw i (3) j=1 with i = 1,...,n. In compact form we can write equation (3) as: Aw = nw (4) with w = (w 1,...,w n ). It is easy to see that (4) is an eigenvalue/eigenvector problem where n is the eigenvalue and w is the associated eigenvector. Owing to the particular form of the matrix A, if we denote with λ i (i = 1,...,n) its eigenvalues, we have the trace of A: n λ i = n (5) i=1 We know from equation (4) that n is an eigenvalue of A so that (from equation (5)) all the other eigenvalues are equal to 0. The case where we know the elements of A through the elements of w is the so called consistent case. In this case matrix A is said consistent 11. If we now suppose to know the elements of the matrix A (through a set of pairwise comparisons) but not the weights w we can solve the problem 12 (4) and obtain the maximum eigenvalue λ max and the associated eigenvector w. We note that if A is consistent, from the preceding 10 As a normalization condition for the vector w = (w 1,..., w n) we have n i=1 w i = We note that in the general case the matrix A is consistent if and only if its elements satisfy conditions 1. and 2. and the following transitivity relation: a ij = a ik a kj (6) with i, j, k = 1,..., n. 12 We note that owing to the structure of a consistent matrix A all the eigenvalues are non negative. In the general case a problem such as: Aw = λw (7) can be solved by imposing det(a λi) = 0 so to define the eigenvalues and the associated eigenvectors. 5

8 remarks, we have that λ max = n is the only non null eigenvalue to which the required eigenvector of the weights is associated. If, on the other hand, A is not fully consistent we have that λ max n and the other eigenvalues are such that λ i 0. In this case the eigenvector 13 w represents a proxy of the real eigenvector w and such approximation is the better the more λ max tends to n. The method has been indeed endowed by Saaty (Saaty (1980)) with a criterion that allows the evaluation of the consistency both of the matrix A and of the eigenvector w we obtain from it. Such criterion, if it is violated, does not prevent the use of such results but simply gives a strong hint that pairwise comparisons must be carefully revised so to attain to a better set of pairwise rankings. The criterion is basically grounded on the definition of a consistency index (C.I.) and a consistency ratio. The former is defined as: C.I. = λ max n n 1 Such index is compared with the average random index that represents the consistency index of a randomly generated reciprocal matrix on the scale Random index allows us to obtain the consistency ratio index as a ratio: consistency ratio = consistency index random index Values of consistency ratio lower than 0.10 define the matrix A we are working with as acceptable, slightly higher values (between 0.10 and 0.20) must be considered with care, really higher values (greater than 0.20) should turn into the rejection of the matrix 15 A. In Saaty (1980) the following table of averages random index values is provided: Table 1: Values of average random index (lower row) as a function of matrix size (upper row) (8) (9) where the values on the first row are the dimension of A whereas those on the second row are the values of the corresponding average 16 random index. At this point we have the matrices A i and the associated eigenvalues λ i and eigenvectors 17 w i and of each matrix we can say if it is enough consistent or 13 Again such a vector must satisfy the normalization condition n i=1 w i = With the expression 1 9 we denote the closed set of integers from 1 to In Saaty and Kearns (1985) at page 34 the authors state that a value of consistency ratio greater than 0.2 should impose a revision of the judgments. 16 It is easy to understand why in cases n = 1 and n = 2 the problem of consistency cannot arise. More precisely the problem of consistency arises only when the dimension of the matrices is greater than In order to avoid any conflict with the i th component w i of a generic vector w we use a bold face notation to denote the eigenvector associated to matrix A i. 6

9 not. We have now to combine all these pieces together in order to obtain a ranking of the alternatives with respect to the main goal. If we call A 1 the matrix of the pairwise comparisons between the n 1 elements at level 1 with respect to the main goal (level 0) we have a vector of the weights of n 1 components that we may call L 1. If at level 2 we have n 2 elements through pairwise comparisons we get n 1 matrices of size n 2 n 2 and therefore n 1 eigenvectors of n 2 elements each. In this way we can construct an n 2 n 1 matrix and call it L 2. At this point if we want to evaluate the weights of the elements at level 2 with respect to the main goal we can simply evaluate the product 18 : L 2 L 1 (10) so to get a normalized vector of n 2 elements. In a similar way we can define the matrices of the pairwise comparisons of the elements at level 3 with respect to those at level 2, be it A 2, and define the matrix L 3 of the vectors of the weights. In order to get the weights of the elements at level 3 with respect to the main goal we can evaluate: L 3 L 2 L 1 (11) so to get a normalized vector of n 3 elements. Further practical details will be given in the next sections. For the moment we only note that through equations such as (10) and (11) we flatten the hierarchy by evaluating the priorities of the elements of any level with respect to the root (or the main goal). The last step is a set of computationally light procedures for the evaluation of the normalized eigenvectors from the matrices A i without solving the associated characteristic equations 19. In Saaty (1980), pages 19 and 20, five methods of increasing precision and complexity are provided. All such methods are based on the particular form of matrix A. 1. The crudest 20. We sum the elements of each row and divide such a value with the sum of all the elements of the matrix. The ratio for the i-th row gives the i-th element of the eigenvector w that is normalized by construction. 2. Better. We sum the elements of each column and then we evaluate the reciprocal of each sum. To normalize we divide each reciprocal with the sum of the reciprocals. 3. Good. We evaluate the sum of the elements of each column and divide each element of a column for that sum (we normalize each column) so to obtain a new matrix. At this point we sum the elements on each row of the new matrix and divide the sum for the dimension of the matrix. In this way we evaluate an average over the normalized columns. 18 To simplify notation we use only row vectors and no special symbol to denote transposition. Depending on the context row vectors must be seen as column vectors so to give the correct meaning to the various operations. 19 The methods we list can be used to find crude estimates of normalized eigenvectors and prove obviously unnecessary whenever exact methods are available. 20 We use Saaty s terminology, as in Saaty (1980), page 19. The term better refers the second method to the first one. 7

10 4. Good. We multiply the elements of each row among themselves, evaluate the n th root (if n is the dimension of the matrix) of that value and, lastly, normalize each of such values. 5. Exact solution. We raise the matrix A to an arbitrarily large power and then divide the sum of the elements of each row of the resulting matrix by the sum of the elements of such matrix. We note that both precision and computational complexity increase from the first method to the last one and that the precision of every method is measured by comparing its results with those obtained by solving the corresponding characteristic equation. We note that if we evaluate the principal eigenvector w we can evaluate the associated eigenvalue by solving directly equation (7). In this way, if the matrix A is consistent, we get n identical values; otherwise we get n slightly different values that we can average to get the true value of λ max to be used to evaluate the degree of consistency of the matrix. 4 A few short notes on electoral systems We present here some short notes and comments on electoral systems (di Cortona et al. (1999), Saari (2001) and Bouyssou et al. (2000) among the many). An electoral system represents a process (di Cortona et al. (1999)) that can be decomposed in the following phases: the definition of the electoral rules, the vote expression, the vote-to-seat translation and the government formation. As such it is a very complex process, nevertheless well suited for a unified formal description with the language of elementary set theory (di Cortona et al. (1999)), and whose performance can be measured with a set of criteria and indicators (di Cortona et al. (1999)). Though complex, an electoral system, starting from each voter s ranking of a set of alternatives (the candidates) from the best to the worse without ties 21, aims at aggregating such rankings in a global social ranking. Unfortunately this is a very hard task and literature is full of impossibility results (see Saari (2001) for many paradoxes and some possible solutions). The main results we want to cite here are 22 : 1. Arrow s [im]possibility Theorem 23 that (Bouyssou et al. (2000)) states that, with more than two candidates, there is no aggregation method that 21 As it will be evident from our examples, in this paper we are going to relax such a hypothesis and use also an indifference relation among the alternatives. 22 We do not aim at giving here a full list of the theoretical negative results of the theory. We only want to mention the most famous sources of troubles. Other results may be found, for instance, in Balinski and Young (1982), as the impossibility of satisfying, at the same time, population monotonicity and staying within the quota in the case of proportional representation methods. 23 Such a theorem comes in three versions (Taylor (2005)): one for the Social Choice Functions, one for the Social Welfare Functions and one for the Voting Rules. Anyway, in all the versions it prevents even a minimal set of properties from being satisfied at the same time. 8

11 can satisfy simultaneously the properties of Universal Domain, Transitivity, Unanimity or Pareto condition (or principle), Binary Independence and Non-dictatorship; 2. Sen s Theorem (Saari (2001)) that is based on a condition of Minimal Liberalism (ML) 24 and states that with more than two alternatives and two or more voters with a Social Welfare Function, if Universal Domain, ML and Pareto are satisfied we are bound to have profiles (or sets of preferences) that have cyclic outcomes (and so fall in the Condorcet paradox of voting); 3. Gibbard-Satterthwaite Theorem (Bouyssou et al. (2000)) that concerns strategic voting (or the convenience of not expressing one s true preferences) and that states that with more than two candidates there exists no aggregation method that satisfies simultaneously the properties of Universal Domain, Non-manipulability and Non-dictatorship. We note that the properties we have listed with Arrow s [im]possibility theorem are really minimal for any real democratic process and that things are even worse (Bouyssou et al. (2000)) if we wished to define a method that satisfied additional properties such as Neutrality, Separability, Monotonicity, Non-manipulability and so on. Similar considerations hold also for Gibbard-Satterthwaite Theorem and Sen s Theorem. Both are hard to accept (this is true also for Arrow s Theorem, see Saari (2001) for a deep discussion and some tentative solutions) and stir up our hope of designing a perfect voting system. Anyway ranking alternatives is needed in many fields so that many imperfect voting systems have been devised and used since a long time. Here we only note that ML possibly should be put in context with a new examination of the Condorcet voting paradox and that the danger of manipulability can be reduced by imposing constraints that make harder the presentation of stray (dummy) candidates/alternatives. 5 The desired properties or the wish lists In this section we start with the wish lists of the electoral systems. Unfortunately such lists, as it should be clear from section 4, are nothing more than unachievable goals. Anyway, our main goal is to recall some basic definitions in order to frame them in the context of the present paper. We start with a first wish list or a first group of basic properties that are involved in Arrow s Theorem. We derive our definitions essentially from Bouyssou et al. (2000) and di Cortona et al. (1999). (1) Universal Domain implies that the chosen aggregation method must be universally applicable so that from any rankings provided by the voters it must yield an overall ranking of the candidates so to rule out methods that would impose some restrictions on the preferences of the 24 A Social Welfare Procedure or Function, or a procedure for the ranking of a set of alternatives, is said to satisfy ML if (Saari (2001)) each of at least two voters is decisive over a pair of alternatives so that his/her ranking of such pair determines that pair s societal ranking. 9

12 voters (Bouyssou et al. (2000), page 17). (2) Transitivity requires that the aggregation of the rankings must be a ranking, with possible ties, that satisfies transitivity. (3) Unanimity or Pareto condition 25 implies that, if each voter ranks a candidate higher than another, this ranking must be reflected in the overall ranking. (4) Binary Independence (or Independence from Irrelevant Alternatives) requires that the relative position of two candidates in the overall ranking depends only on their relative position in each voter s ranking so that all the other alternatives are seen as irrelevant with regard to those candidates. (5) Non-dictatorship means that there is no voter that can impose his/her ranking as the overall social ranking. We note that Condorcet method 26 satisfies properties (1), (3), (4) and (5) so that, by Arrow s Theorem, it must fail property (2) (and indeed a Condorcet winner does not necessarily exist owing to the existence of cycles among candidates) whereas Borda method 27 satisfies properties (1), (2), (3) and (5) so that, by Arrow s Theorem, it must fail property (4) and indeed Borda method suffers from this drawback that can be exploited to manipulate the overall ranking 28. We can now enlarge the basic list (and make things even worse) by adding the following properties 29. (6) Anonymity (Taylor (2005)) requires that voters are treated the same way so that the overall ranking is independent from any permutation of the voters. (7) Neutrality (Taylor (2005)) means that alternatives are treated the same way so that the overall ranking is independent from any permutation of the alternatives. (8) Separability (Bouyssou et al. (2000)) requires that if we perform an election with two separate sets of voters and obtain a winner candidate on each set such candidate remains a winner if we repeat the election with the same method on the union of the two sets of voters. (9) Monotonicity (Bouyssou et al. (2000), page 11) requires that an improvement of a candidate s position in some of the voter s preferences cannot lead to a deterioration of his position after the aggregation. (10) Non-manipulability is a very complex issue (Taylor (2005)) but essentially it means that the overall ranking of a set of candidates does not depend either on the agenda or on the presence of stray candidates or on the expression of non true preferences. In order to deepen our examination of electoral systems, we can give the wish lists of properties also for both majoritarian methods (where only one seat is assigned in every district) and proportional methods (where S seats are assigned in every district), di Cortona et al. (1999). Majoritarian methods are characterized by the following properties. (CW) Condorcet winner: it is the winner of all pairwise comparisons, if it exists it should be the winner of the electoral competition. (CL) Condorcet loser: a 25 We note that though Taylor (2005) differentiates the two properties we consider them as synonyms. 26 The Condorcet method uses pairwise comparisons between candidates. 27 The Borda method uses a global ranking of the candidates. 28 We note however that Saari (2001), at page 148, states that when ways to circumvent the difficulties of Arrow s Theorem are examined...only the Borda Count survives all of the different requirements. 29 We underline how some of these properties may take a different meaning when we will examine proportional and majoritarian methods. 10

13 method should not choose the candidate that loses every pairwise comparison with all the other candidates. (M) Monotonicity: a method is monotone if the number of seats assigned to a party does not decrease if the number of its supporters grows. (PP) Pareto principle: if all the voters prefer a candidate to another the latter cannot be chosen. (WARP) Weak Axiom of Revealed Preference: it requires that, (a), if a candidate is a winner on a set X of candidates it must remain a winner also on any subset X X to which he/she belongs and that, (b), if there are ties among candidates in X X those candidates at par must be all either included or excluded from the final set of winners in X. This axiom is used to get voting methods immune from manipulations on the set of candidates through the addition of stray candidates. (PI) Path independence: a method satisfies path independence if the outcome is independent from the ordering of the phases that are used for the selection of the candidates. We note that First-past-the-post 30 method satisfies (6), (PP) and (WARP) whereas Double ballot and Single transferable vote 31 methods satisfy (6), (CL) and (PP). Proportional methods are characterized by the following properties. (HM) House monotonicity means that if the number of seats passes from S to S+1 no party gets fewer seats. (QS) Quota satisfaction requires that the number of seats each party receives is as close as possible to its exact quota and so to a percentage of the total seats that is almost equal to the percentage of the votes it receives. (PM) Population monotonicity (di Cortona et al. (1999)) if a party (or state) with a growing weight cannot lose a seat in favor of a party (or state) with a declining weight. (C) Consistency requires that any partial assignment is itself proportional. (S) Stability means that whenever two parties merge in a coalition (or a new party) they do not get fewer seats that those they get as separate entities. We note that the Quota method 32 satisfies (6), (HM), (QS), (C) (only with 30 This is a method of the majoritarian family where winner takes all so that in every uninominal district the candidate that receives a majority of votes is elected independently from the obtained percentage. Double ballot is articulated into two rounds (with or without threshold) so that the second occurs only if in the first no candidate gets an absolute majority of votes. If this occurs the most voted candidate is elected. In the case of Single transferable vote we have an iterative procedure where the less voted candidate is dropped and his votes are transferred to his next most preferred candidate still in competition until when one candidate reaches more that 50% of the votes cast. 31 In di Cortona et al. (1999), page 29, Single transferable vote is cited among proportional methods owing to its use of Droop quota and for its proportional redistribution of excess votes from elected candidates to second choice ones whereas in the Table at page 78 it is put in comparison with other most popular majoritarian methods. In the present paper we chose the latter classification in order to use that Table for our comparisons of majoritarian methods, see section Quota method assigns the seats by evaluating a quota and rounding it up or down. Such a quota is evaluated as q i = S V v i where S is the total number of seats, V is the total number of votes and v i is the number of votes of party i. Divisor methods (really a whole class of methods) are based upon a set of divisors. In this case, for each party and each divisor, we compute the ratio of the number of votes received by a party and that divisor so that the S seats are assigned to the parties that achieve the S largest such ratios (di Cortona 11

14 regard to pairs of eligible parties) and (5), Divisor methods satisfy (6), (HM), (P M), (C) and (S) (only in particular cases) whereas Largest remainders methods satisfy (6), (QS) and (S). 6 A simple ranking Let us start with a very simple example. We say this example is simple since voters have to rank concrete alternatives rather that abstract properties as they do in section 7. We suppose to have three voters (See note 3) and four alternatives that must be ranked so to define the best alternative among the four or, at least, a total ordering on them with possible ties. In this case alternatives may represent either candidates in an electoral competition or even different kinds of pizza flavors. Figure 2: Three voters and four alternatives The situation is shown in Figure 2. The Main Goal (i.e. the ranking of the alternatives) is labeled as MG whereas voters are labeled as v1, v2 and v3 and a similar convention holds also for the four alternatives. As a first step we evaluate the normalized vector of the weights of the voters with regard to MG. It is easy to see that imposing a full symmetry of the three voters we get a fully consistent 3 3 matrix (with all elements equal to 1) to which it corresponds the eigenvalue λ max = 3 and an eigenvector L 1 = (1/3, 1/3, 1/3). This result is consistent also with our intuition of a fair evaluation tool since it seems obvious that the three voters have the same weight in the process. As the successive (and last in this case) step we have to evaluate three 4 4 matrices of the pairwise comparisons of the four alternatives, each matrix with regard to a single voter. For these evaluations we use the scale (See note 14.) 1 9 we introduced in section 3 and suppose that the three voters have respectively the following preferences on the alternatives 33 : et al. (1999), page 26). Largest remainder methods use such a quota q i to evaluate a remainder r i = q i q i to assign the not yet directly assigned seats to the parties with the highest values of the remainder. 33 We denote with > a binary relation of strict preference and with a binary relation of indifference. No rationality hypothesis is imposed on the voters and such relations are supposed to be endowed with classical properties. 12

15 1. a1 > a2 > a3 > a4, 2. a4 > a3 > a1 > a2, 3. a3 a4 > a2 a1. The three matrices are therefore those in Tables 2, 3 and 4. Such matrices 34 are evaluated as follows: voters are requested to assign to every pair of alternatives a numerical evaluation on the scale 1 9 so to satisfy the requirements of Analytic Hierarchy Process and, at the same time, their preference orderings. The aim is the definition of positive reciprocal matrices that are consistent at the most and reflect voters judgments about the various alternatives (see e Costa and Vansnick (16 June 2008) for a criticism about this point). v1 a1 a2 a3 a4 a a2 1/ a3 1/5 1/2 1 2 a4 1/7 1/3 1/2 1 Table 2: Pairwise comparisons with regard to v1 v2 a1 a2 a3 a4 a /2 1/4 a2 1/2 1 1/3 1/6 a /3 a Table 3: Pairwise comparisons with regard to v2 v3 a1 a2 a3 a4 a /5 1/5 a /5 1/5 a a Table 4: Pairwise comparisons with regard to v3 By using the method of the n th root of the product it is easy to evaluate the eigenvectors of such matrices, the corresponding eigenvalues and verify that every consistency ratio falls below the threshold suggested by Saaty so that each matrix is consistent. Such eigenvectors form the matrix L 2 of Table 5. At this point the normalized vector of the weights of the alternatives with respect to the main goal can be easily evaluated as: 34 This holds also for the matrices we use in sections 7 and 8. w = L 2 L 1 (12) 13

16 L 2 = w 1 w 2 w Table 5: Matrix L 2 of the eigenvectors alternatives versus voters so to get: w = ( ) (13) From expression (13) we can easily deduce the following (and possibly counter intuitive) ordering on the alternatives: a4 > a3 > a1 > a2 (14) At this point a question (at least) spontaneously arises: what did we get and for what aim? We got a ranking, right. Can we use it as if it was an election outcome? Maybe. The main problem to face is the inconsistency issue. In the general case, indeed, we can have one or more inconsistent matrices: how can we deal with this? There is any threshold above which we should reject a ranking? Or should we consider it anyhow valid? Some of these questions will remain unanswered in this paper, owing to their complexity, others will find at least partial answers in section 9. 7 Some more complex rankings After the simple example of section 6, in this section we have major aims. We think that the cases we deal with in this section are more complex than the example we presented in section 6 not from a computational point of view but essentially because in these cases we deal with abstract properties of electoral systems rather than with concrete alternatives. We use a set of properties that characterize the families of majoritarian and proportional methods to obtain a ranking of those properties and, depending of this ranking, define the perfect method within each family. The first very simplified situation is illustrated in Figure 3 where we suppose to have (only) four voters who rank the six main properties that characterize proportional methods: (6) Anonymity (A), (HM) House Monotonicity, (QS) Quota Satisfaction, (P M) Population Monotonicity, (C) Consistency and (S) Stability. If we perform the pairwise rankings for each voter we can get the Tables 6, 7, 8 and 9. We assume that the four voters have the following preference orderings respectively: 1. A > HM > QS > PM > C > S 14

17 Figure 3: Ranking properties of proportional methods 2. A HM > QS PM > C > S 3. S > C PM > A > HM QS 4. QS > HM A > PM C S Also in this case, by using the method of the n th root of the product, it is easy to evaluate the eigenvectors of such matrices, the corresponding eigenvalues and verify that every consistency ratio falls below the threshold suggested by Saaty so that each matrix is consistent. Such eigenvalues form the matrix L 2 of Table 10. Also in this case, if we suppose that the four voters have the same weight with regard to the Main Goal (MG) and define the proper matrix at level 1, we get a matrix whose elements are all 1. Both by performing the calculations and by using fairness considerations it is easy to see that as eigenvector of level 1 we get L 1 = ( ) so that as L 2 L 1 we get the content of Table 11. In that table the first column contains the vector of the rankings of the properties with regard to the Main Goal, the second contains the listing of the mnemonics of the properties and the last their place in the classification. A close inspection of Table 11 and a comparison with the results of the table at page 83 of di Cortona et al. (1999) allow us to assert that on the basis of our results the best proportional method (or, more correctly, the method our four voters prefer) is the Quota method: from that table we have that Quota method satisfies A, HM, QS, C (but only in special cases) and S. It is obvious that by changing the preferences of the voters (and also by augmenting their number) we surely get a different ordering, probably with ties, to which, again with regard to the cited table of di Cortona et al. (1999), can correspond either another winner or a pair of winners or no winner at all. Let us suppose, for instance, that only voter v2 changes his/her pairwise comparisons and takes those of Table 12 (associated to the following preference ordering PM > A C > S > HM > QS). 15

18 v1 A HM QS PM C S A HM QS PM C S Table 6: Pairwise comparisons with regard to v1 v2 A HM QS PM C S A HM QS PM C S Table 7: Pairwise comparisons with regard to v2 v3 A HM QS PM C S A HM QS PM C S Table 8: Pairwise comparisons with regard to v3 v4 A HM QS PM C S A HM QS PM C S Table 9: Pairwise comparisons with regard to v4 16

19 L 2 = w 1 w 2 w 3 w Table 10: Matrix L 2 of the eigenvectors alternatives versus voters L 2 L 1 = w 0.25 A HM QS PM C S 4 Table 11: Final ranking w of the alternatives and their classification We note indeed how the properties A, HM and QS summed up have a weight of 0.65 or: w A + w HM + w QS = 0.65 (15) If we evaluate the new eigenvector w 2 we get, obviously, a different vector that v2 A HM QS PM C S A HM QS PM C S Table 12: New pairwise comparisons with regard to v2 gives rise to a different second column of matrix L 2. This turns into a somewhat different final priority vector, see Table 13. A close inspection of Table 13 and a comparison with the results of the table at page 83 of di Cortona et al. (1999) allow us to assert that on the basis of our results the best proportional method (or, more correctly, the method our four voters prefer) is, in this case, the Largest remainder methods: from that table we have indeed that Largest remainders methods satisfy A, QS whereas S; Divisor methods 17

20 L 2 L 1 = w 0.23 A HM QS PM C S 2 Table 13: Another final ranking w and classification of the alternatives satisfy A, HM, P M, C and S (but only in special cases) and Quota method satisfies A, HM, QS, C (but only in special cases) and S. We note moreover that the properties A, QS and S count for almost 60% over the total of the six properties and that a change of one voter s opinion from four can be seen as a change of the opinion of the 25% of the voters. Figure 4: Ranking properties of majoritarian methods Now we go through a similar example but with regard to the properties of majoritarian methods. In Figure 4 we again suppose to have (only) four voters who, in this case, rank the six main properties that characterize majoritarian methods and so: (6) Anonymity (A), (CW) Condorcet Winner, (CL) Condorcet Loser, (PP) Pareto Principle, (WARP) Weak Axiom of Revealed Preferences and (P I) Path Independence (P I). In this case we give only the matrix L 2 of the eigenvectors and the vector of the priorities of the properties with regard to the main goal (and make some comments). We note that the four matrices that provide the eigenvectors of L 2 are based on the following preference orderings: 1. A > CW > CL > PP > WARP > PI 2. CW CL > PP > PI > A WARP 3. PP > A > PI > WARP > CW CL 4. PP > PI > A WARP > CW > CL 18

21 w 1 w 2 w 3 w 4 w 0 w A CW CL PP WARP PI Table 14: Majoritarian methods: the eigenvectors of the weights w i and the final vector of the weights w. The four columns on the left are the columns of the matrix L 2 of the eigenvectors of the alternatives as to the voters. The fifth column contains the eigenvector L 1 of the weights of the voters as to the Main Goal whereas the sixth column represents the vectors of the weights of the alternatives as to the Main Goal. The four matrices can be shown to be fully consistent according to the Saaty criterion. The final results, that we show in compact form to save space, are those of Table 14. The sixth column of table 14 is obtained by a matrix and vector multiplication between the first four columns and the fifth column. From the values of the sixth column we can devise the following ordering: with: and: PP > A > CW > CL > PI > WARP (16) w PP + w A + w CL = 0.64 (17) w PP + w A + w WARP = 0.57 (18) Result (16), in the light of the table at page 78 of di Cortona et al. (1999) and equations (17) and (18), can be a little bit difficult to interpret. From that table we indeed have that: First-past-the-post method satisfies A, P P and W ARP; Double ballot and Single transferable vote methods satisfy A, CL and PP; Approval voting method satisfies A, WARP and PI. By confronting all such informations we can say that: 1. both Double ballot and Single transferable vote methods satisfy A, CL and PP (and only these properties); 2. only First-past-the-post method satisfies A, P P and W ARP. By adding the corresponding elements of the priority vector up, all we can devise therefore is the following preference ordering: Single transferable vote Double ballot > First past the post (19) and reach a final decision by using other criteria. 19

22 8 Two other attempts of ranking In this section we show two other attempts of performing a ranking of electoral systems. In the first simple example we consider high level properties such as (1) Universal Domain (UD), (2) Transitivity (TR), (3) Pareto condition (P) and (4) Binary Independence (BI). Figure 5: Ranking electoral systems through ranking some basic properties Figure 5 shows the case of four voters that perform a ranking (trough pairwise comparisons) of these basic properties. The four matrices of the pairwise comparisons are those of Tables 15, 16, 17 and 18. Such Tables are respectively based on the following preference orderings of the four voters: 1. TR > UD BI P 2. P > TR > UD > BI 3. TR > P UD > BI 4. UD > P > TR > BI By performing the proper calculations it is easy to verify that all such matrices are consistent and that the eigenvectors 35 are those of the first four columns of Table 19 whereas the fifth column represents the eigenvector of the matrix of the pairwise comparisons of the four voters with regard to the main goal. From both calculations and fairness considerations it is easy to see that such a vector has all components equal to The sixth column gives the global weights or priorities of the four properties with regard to the main goal. With our data we have T R P > U D > BI. Such a ranking is satisfied, for instance, by the Borda count that does not satisfy binary independence. This does not mean of course that Borda count is the only method that satisfies our data but only that it is one of the methods that do that and, so, can be legitimately chosen. We note indeed that: w TR + w UD + w P = 0.90 (20) so that Binary independence can be surely neglected. 35 We note that in all the following cases all eigenvectors are evaluated according to the method of the n th root of the product and all eigenvectors are normalized. 20

23 v1 TR UD BI P TR UD BI P Table 15: A first ranking of electoral systems, case of v1 v2 TR UD BI P TR UD BI P Table 16: A first ranking of electoral systems, case of v2 v3 TR UD BI P TR UD BI P Table 17: A first ranking of electoral systems, case of v3 v4 TR UD BI P TR UD BI P Table 18: A first ranking of electoral systems, case of v4 The second attempt involves a ranking between majoritarian methods M and proportional methods P if we consider them as two opposing families of methods. The basic idea is shown in Figure 6. In this case we suppose to have three voters (labeled as v1, v2 and v3) that use four properties (labeled as p1, p2, p3 and p4) to obtain a ranking between the family of majoritarian methods M and that of proportional methods P to see which family is better on the basis of the pairwise rankings of such properties 36. In this case we have a rooted hierarchy where the leaves are at level 3 so we have to define the matrices for three layers and precisely: four 2 2 matrices at level 3 to which there corresponds one 2 4 matrix of four eigenvectors L 3 ; three 4 4 matrices at level 2 to which there corresponds one 4 3 matrix of three eigenvectors L 2 ; one 3 3 matrix at level 1 to which there corresponds one 3 1 matrix of one eigenvector L In this case we have two alternatives (M and P) that the voters must rank according to four properties p i (to be specified) with regard to the Main Goal of choosing one of the two alternatives. 21

24 w 1 w 2 w 3 w 4 w 0 w TR UD BI P Table 19: The eigenvectors of the weights w i and the final vector of the weights w. The four columns on the left are the columns of the matrix L 2 of the eigenvectors of the alternatives as to the voters. The fifth column contains the eigenvector L 1 of the weights of the voters as to the Main Goal whereas the sixth column represents the vectors of the weights of the alternatives as to the Main Goal. Figure 6: Majoritarian or proportional? The basic dilemma In this way we can evaluate the priorities vector w of the two alternatives M and P with respect to the root of the hierarchy (or the main goal) as a product of matrices: w = L 3 L 2 L 1 (21) From considerations we have already made elsewhere in sections 7 and 8 of this paper it is easy to see that L 1 = ( ). The hard part is the definition of the four properties. We can try with the followings properties (corresponding respectively to the pi of Figure 6) 37 : (11) Electoral Participation (EP) defined as the ratio between the number of vote cast and the difference between the total number of voters and the number of vote cast; (12) Number of Political Parties (NPP) defined through parameters that count both the number of parties that compete in a given election and their relative strength; (13) Electoral Volatility (EV ) as a measure of the electoral fluxes among the 37 We give only informal definitions of such properties. For further and more exact details see di Cortona et al. (1999). 22